Abstract: The transparency of NSFW AI (AI systems meant to detect and filter out Not Safe For Work content) is an important consideration in both the research community building these technologies as well as entities deploying it. The above sets the bar high for what constitutes "reasonable transparency" on how exactly an AI's processes, its criteria of decision-making and data handling practises are communicated to users & stakeholders. Transparency especially is key for NSFW AI, because of how high impact (and sensitive) the content they handle are.
Is NSFW AI transparent? The answer is complicated. There is still work to do in the realm of transparency, even if we have made some progress! A significant worry is that we do not know what data were the models trained on. This NSFW AI provides datasets that can contain up to millions of explicit images, videos and text pulled down from the internet as part of their training sets but with its origin/consents often unrevealed. For example, a 2021 study found more than 60% of AI developers fail to provide full details on the sources for their training data — indicating serious ethical and legal breaches.
Similarly, the opacity of decision-making algorithms is also at stake. The employed AI models for NSFW classification usually consist of complex machine learning systems—like a convolutional neural network (CNN) which function as black boxes and provide results without providing transparent reasons behind the decisions. The fact that there is little explainability. Results in mistrust on the user's side, especially when their content get flagged or removed without a clear reason of why it happened. Creators on platforms such as OnlyFans or Patreon, similarly hate not knowing how to improve their success rate and are therefore frustrated when content is being taken down by AI without the super-necessary but also awfully useless factor of feedback – again in a non-transparent manner.
Plus, there is bias in NSFW AI as well, that adds to the transparency issues. So, for example if the AI is trained with biased data in may start to flag more posts from a particular demographic or cultural context and hence this might lead to an unfair treatment. The first known incident of Tumblr's NSFW AI dealing differently with content depicting different skin colors in 2020 and the subsequent 'algorithm seems racist' debates. This makes it difficult to find and repair biases in the model without transparency into how the AI is being trained, tested or making decisions.
Work on increased transparency continues. To help some set of developers have started the AI explainability tools, to make it a little more understandable for other people making use of this decision. Google is also exploring explainability with its Explainable AI project to show how an ai model comes to a decision so it can provide content moderation in searches, and parents would have more idea of why some image gets flagged as NSFW. But these tools are so far primitive and not commonly used in the NSFW AI area.
This also extends to user privacy and how data is handled. Most users are not in the know of where their data come and how NSFW AI Systems store, gather or utilize these information. Since the information being covered is itself so sensitive, it should go without saying that transparency about data practices is essential. In Europe, The General Data Protection Regulation (GDPR), for example requires an explicit disclosure of algorithms and processing activities but enforcement in the case with NSFW AI has been mixed.Slf
As Tim Berners-Lee, the creator of The Internet and World Wide Web said “We need to re-decentralize the web(so everyone can participate in our growing digital world), by giving individuals greater control over how their data is collected used.” This line hammers home the point of openness and transparency in digital and AI technologies.
As you can see, even though some initial attempts at making NSFW AI more transparent have been made, in practice there appear to be quite insurmountable challenges. Algorithms are so complex, decision-making processes that seem rather opaque and where issues of data privacy or bias in decisions (sic) give transparency nothing more than a pale reflection.
To read more about NSFW AI and other related uses, you can visit here as a nsfw ai resource.